'It's an understatement to say AI will be a game changer for global health': Dr. Zameer Brey

Dr. Zameer Brey envisions a future where every frontline medical worker will be able to hold an AI-enable smartphone to a patient's chest and diagnose a heart problem.

"I'm not talking about replacing the cardiologist," he says. "I would ping the cardiologist and say, 'Hey, my tool just suggested with 90% accuracy that this is a left bundle branch block with some SD segment elevation. I probably should refer the patient to you.'"

It's scenarios like this that keep Dr. Brey, who leads technology diffusion for the Gates Foundation, awake at night with possibility. "Quite honestly, the thing that gets me super excited and sleepless right now is that we can we upskill health care workers to do tasks that otherwise would've taken them years of training or inconvenienced patients with week-long or month-long appointments at distant facilities," he says.

In this second part of our interview with Dr. Brey, Exemplars News asked him about how AI will transform global health, as well as the potential risks and benefits of the technology, especially for LMICs in the Global South.

Broadly, how do you think AI will impact global health over the coming decades? What, in your mind, are the potential risks and benefits?

Dr Brey: It's an understatement to say AI will be a game changer for global health. I'm not saying that because I want to be overly optimistic. I'm saying that because of what I see already happening with the almost 50 Grand Challenges projects we funded directly, what I'm seeing some of these grantees continuing to do, and the technological capabilities of AI increasing exponentially.

Let me give you an example. A baby presents in respiratory distress to a doctor in rural Sudan. They ask ChatGPT what to do. The initial response actually is pretty spot on. 'We want you to do the ABCs, and then we want you to specifically ask about X, Y, and Z.' But if this baby is in distress, that's certainly not going to be enough. ChatGPT can ask the doctor, 'What do you have in your consulting room?' Let's say it's a newly qualified doctor, so they don't know. They can take a picture of the consulting room. ChatGPT identifies a CPAP machine in one corner of the room and says, 'This is a CPAP machine. I'm going to coach you through how to apply this machine in this situation. These are the six steps. Then we're going to make sure that it's working. This is what you're going to do to monitor. If this baby continues to desaturate, this is who you're going to contact.'

I've personally been experimenting with the tools. I took a picture of an X-ray. Not even a real case. I just said, 'What's wrong?' and it got the diagnosis spot on. It was a right-sided neutral effusion. Then it gave me the differential diagnosis. Told me how to investigate this patient. Can we do the same for ECGs, EEGs, CT scans, and heart sounds? Some cardiologists will say, 'You can't pick up an S3 ventricular gallop in heart failure until you've listened to 10,000 heart sounds.' That may be true. But in LMICs there are only a few cardiologists for parts of the population. What if we could put a phone on someone's chest and record it and have it interpreted?

There are some risks here and we in the development community have a responsibility to make sure [these technologies are equitably distributed]. Even more so, the Gates Foundation has a responsibility to get these technologies into the hands of the people who will benefit the most. I think there's a risk that the way technology gets developed, designed, and innovated is to serve the needs of the Global North because that's where a lot of the capital is flowing. There's a risk that the people who stand to benefit the most [from AI in health] will ultimately not see the benefits or they will trickle down over many, many years. We've got to fundamentally change that narrative. The role of the Gates Foundation is to try and create a space for scientists, technologists, and researchers in LMICs to really leverage AI, demonstrate the use cases, and support scaling up.

I will say that one other risk is making sure we continue to use context-specific data for all the large-language models. If we don't do that, it will be garbage in, garbage out. It presents really risky kinds of things. I recently gave one of the models a very typical TB-coinfected HIV case. The model really struggled and failed fundamentally. If it was a first-year medical student, it would not have passed. It failed on the differential diagnosis, on the diagnostic paradigm, on the treatment path, etc. The moment I added my location – I added just one word to the prompt – it got 85% of the diagnosis on the first try. That just underscores how important context and data are. Basically, we need more data.

Another risk is bias. Let me give you another example. I asked a model to create an image of a community health care worker in Africa. It gave me an image of a white male community health care worker wearing gloves. I've never seen a white male community health care worker in Africa. Then I asked it to give me an image of the world's top decision makers doing a field visit in Africa. In the image, they were all wearing dress shirts, blue blazers, and ties. The hard challenge is that the bias is real. I don't think that should be a reason why we slow down, but I think it should be something we pay attention to as we build these technologies.

You mentioned the Gates Foundation funded almost 50 projects developing AI, global health, and development solutions for communities. Tell us more about that.

Dr. Brey: I think it was one of the most prolific Grand Challenges in the foundation's history. Not because we made almost 50 grants, but because it dispelled a myth. That myth was that LMICs are not ready for AI. We received 1,334 applications in 14 days. I think that's a very clear expression of how ready scientists and technologists are in the Global South to take on some big, complex challenges. Ultimately, we awarded nearly 50 grants, but there were at least 100 really high-quality proposals and some of our other funders are thinking about how to take on some of them.

How can we ensure AI technologies for global health are accessible to underserved populations and LMICs?

Dr. Brey: Language is a really big one. There are 2,000 languages globally, but we see very little language representation [in AI models]. One of the biggest barriers to ensuring these technologies are accessible is making sure we have a diversity of languages that folks will engage with. Of course, this is important in terms of accessing the technologies. But it's even more important when literacy levels are low because we can't always expect people to type [their prompts] in a local language if we want solutions to be rolled out to over a billion people. We have to make sure they can use voice-based technologies. Therefore, language has become even more important than it was six months ago.

The second is compute capacity, which is related to being able to do research and scale up AI technologies in LMICs. I'm really disappointed how this is currently playing out. In many ways it's the same storyline as COVID vaccines – who can pay the most, who can get the most and who can get them the fastest. Many good [AI] researchers and scientists want to do cool things, but they've been told, 'There's a long line so good luck.' Several researchers have come to us pleading for access to compute. We know they're doing substantive work, but they just don't have access to compute, let alone high- performance chips. They can't even access cloud compute because of structural issues like credit cards, forex flows, service providers not being interested in their markets, etc. One of the things we're thinking about is a GAVI for compute. It may be the time for us to say, as a global community, that we recognize the importance AI will play in development and, therefore, to unlock that we've got to come up with a preferential procurement mechanism that lowers the price of compute and distributes it more equally.

How can the biases in AI be prevented from exacerbating existing inequalities in global health?

Dr. Brey: In some ways this is already playing out since the [AI] tools have been informed with the corpus of data that represents the Global North. Unless we can fix the corpus of data, unless we can use data from the Global South and LMICs to inform the AI models, they will be less relevant in our contexts.

What in your upcoming work is most exciting or energizing for you right now?

Dr. Brey: I think we're on the verge of a transformation in the global health space that none of us are actually ready for. Quite honestly, the thing that gets me super excited and sleepless right now is that we can we upskill health care workers to do tasks that otherwise would've taken them years of training or inconvenienced patients with week-long or month-long appointments at distant facilities. That's the opportunity and I think we're pretty close to it. Imagine if I, as just a plain old GP, could do a cardiology consult. I'm not talking about replacing the cardiologist. I would ping the cardiologist and say, 'Hey, my tool just suggested with 90% accuracy that this is a left bundle branch block with some SD segment elevation. I probably should refer the patient to you.' That saves the cardiologist time. We're not there yet, but this is probably why we shouldn't be sleeping.